[Corpora-List] Considering Distributions Across Texts

Marc Brysbaert marc.brysbaert at ugent.be
Mon Mar 3 13:06:49 CET 2014


James Adelman (University of Warwick) has argued that contextual diversity (number of texts in which a word appears) is better than word frequency (number of times a word occurs). You find the paper here:

http://homepages.warwick.ac.uk/~pssgar/cd/cdps4.pdf

Also in our work, we find that contextual diversity predicts word recognition times better than word frequency.

This being said, most of the difference is due to the nonlinear frequency function and to the fact that proper nouns seem to be unevenly distributed across texts. When you take these two factors into account, there is not much difference any more. In an unpublished ms, Emmanuel Keuleers found that you find a contextual diversity ‘advantage’ if you randomly distribute the words over the files. This agrees with the fact that much of the superiority is due to the nonlinear function of frequency (there is a floor effect for word frequency above 50 per million words; this tail is much shorter in contextual diversities than in word frequencies).

So, in summary you can use both contextual diversity and word frequency.

Best, mb

From: corpora-bounces at uib.no [mailto:corpora-bounces at uib.no] On Behalf Of Adam Kilgarriff Sent: maandag 3 maart 2014 12:40 To: Don Tuggener Cc: corpora at uib.no Subject: Re: [Corpora-List] Considering Distributions Across Texts

Dear Brian,

Are the 300-400 texts from 300-400 different people? If yes, then, if you use document frequencies ("how many documents does this word/construction/... occur in") rather than "how many times does it occur" you will cancel out skews based on particular people.

If the texts are all the result of the same essay question, or a limited number of essay questions, then of course you have the bias related to what the students were being asked to write about.

I'm a sceptic about statistical significance testing (for the full argument see here <http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.100.6901&rep=rep1& type=pdf> ) - the main thing is to have a good understanding of the structure of your sample, and the ways that is likely to introduce bias

Adam

On 3 March 2014 11:02, Don Tuggener <tuggener at cl.uzh.ch> wrote:

Hi Brian,

I'm guessing you're looking for tests that help you identify statistical significance of your query results? A good starting point may be: 2010f. Gries, Stefan Th. Useful statistics for corpus linguistics. In Aquilino Sánchez & Moisés Almela (eds.), A mosaic of corpus linguistics: selected approaches, 269-291. Frankfurt am Main: Peter Lang. (http://www.linguistics.ucsb.edu/faculty/stgries/research/overview-research. html)

Best, Don

On Mon, 03 Mar 2014 11:28:35 +0100 corpora-request at uib.no wrote:


> Message: 3
> Date: Fri, 28 Feb 2014 11:16:11 -0500
> From: Brian Schanding <bschanding at gmail.com>
> Subject: [Corpora-List] Considering Distributions Across Texts
> To: corpora at uib.no


>
> Hello,
>
> I'm working on research with learner corpora. My corpora aren't that big
> (approx. 250,000 wds with about 300-400 text files). I wonder what
> research/textbook sources anyone can point me to that discuss the
> importance of considering how many texts in the corpus a language feature
> occurs in (as opposed to merely considering overall frequency of a
language
> feature within a corpus).
>
> Many Thanks!
> Brian

_______________________________________________ UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora Corpora mailing list Corpora at uib.no http://mailman.uib.no/listinfo/corpora

-- ======================================== Adam Kilgarriff <http://www.kilgarriff.co.uk/> adam at lexmasterclass.com Director Lexical Computing Ltd <http://www.sketchengine.co.uk/> Visiting Research Fellow University of Leeds <http://leeds.ac.uk>

Corpora for all with the Sketch Engine <http://www.sketchengine.co.uk>

DANTE: <http://www.webdante.com> a lexical database for English

========================================

-------------- next part -------------- A non-text attachment was scrubbed... Name: not available Type: text/html Size: 10242 bytes Desc: not available URL: <https://mailman.uib.no/public/corpora/attachments/20140303/5de7c5c4/attachment.txt>



More information about the Corpora mailing list